MXNet is a deep learning framework designed for both efficiency and flexibility. It allows you to mix the flavours of symbolic programming and imperative programming together to maximize the efficiency and your productivity. In its core, a dynamic dependency scheduler that automatically parallelizes both symbolic and imperative operations on the fly. A graph optimization layer is build on top, which makes symbolic execution fast and memory efficient. The library is portable and lightweight, and is ready scales to multiple GPUs, and multiple machines.
MXNet is also more than a deep learning project. It is also a collection of blue prints and guidelines for building deep learning system, and interesting insights of DL systems for hackers.
- Design Note: Design Efficient Deep Learning Data Loading Module
- MXNet on Mobile Device
- Distributed Training
- Guide to Creating New Operators (Layers)
- Amalgamation and Go Binding for Predictors
- Training Deep Net on 14 Million Images on A Single Machine
- Documentation and Tutorials
- Open Source Design Notes
- Code Examples
- Pretrained Models
- Contribute to MXNet
- Frequent Asked Questions
- Open sourced design note on useful insights that can re-used by general DL projects.
- Flexible configuration, for arbitrary computation graph.
- Mix and Maximize good flavours of programming to maximize flexibility and efficiency.
- Lightweight, memory efficient and portable to smart devices.
- Scales up to multi GPUs and distributed setting with auto parallelism.
- Support python, R, C++, Julia,
- Cloud friendly, and directly compatible with S3, HDFS, AZure
- For reporting bugs please use the mxnet/issues page.
© Contributors, 2015. Licensed under an Apache-2.0 license.
Tianqi Chen, Mu Li, Yutian Li, Min Lin, Naiyan Wang, Minjie Wang, Tianjun Xiao, Bing Xu, Chiyuan Zhang, and Zheng Zhang. MXNet: A Flexible and Efficient Machine Learning Library for Heterogeneous Distributed Systems. In Neural Information Processing Systems, Workshop on Machine Learning Systems, 2016
MXNet is initiated and designed in collaboration by authors from cxxnet, minerva and purine2. The project reflects what we have learnt from the past projects. It combines important flavour of the existing projects, being efficient, flexible and memory efficient.